News & Updates
The latest news and updates from companies in the WLTH portfolio.
Anthropic researcher resigns, warns AI could spiral out of control as soon as next year
Jacob Coxon's departure marks the latest in a pattern of safety-focused exits from leading AI labs, with colleagues publicly backing his dire assessment A 27-year-old British researcher who worked on pretraining at Anthropic has publicly resigned, warning that the race to build recursive self-improving AI could lead to human extinction before the decade is out. Jacob Coxon walked away from one of the most prominent AI safety companies in the world around September 8-9, and his parting message was not a polite LinkedIn post about "exciting new chapters." It was a warning that the people building these systems genuinely believe they could produce catastrophic outcomes, and that the timeline for things spiraling out of control could be as soon as next year. The insider alarm What makes Coxon's resignation particularly difficult to dismiss is what happened next. His own colleagues at Anthropic started publicly agreeing with him. Evan Hubinger, who holds the title of Alignment Science Lead at Anthropic, backed up Coxon's claims on social media. His personal estimate: a greater than 10% chance of AI leading to massive human fatalities within the next decade. Hubinger went further, acknowledging that Anthropic lacks a clear plan for ensuring alignment with superintelligent systems. Samuel Marks, another Anthropic researcher, echoed similar concerns, suggesting that senior employees inside the company are particularly worried about extinction-level risks. Coxon specifically criticized the use of terms like "crunchtime" and "endgame" within AI development circles. These words frame the pursuit of superintelligence as something urgent and inevitable, a competition to be won rather than a risk to be managed. That framing, he argued, is part of the problem. A pattern of exits Coxon isn't the first safety-focused researcher to leave Anthropic with concerns about the company's direction. Mrinank Sharma, who led the company's Safeguards Research team, resigned in February 2026. Sharma pointed to escalating dangers posed by AI and bioweapons, citing external pressures that he said were compromising the company's ethical conduct. The irony is thick. Anthropic was founded by former OpenAI staff, including CEO Dario Amodei, specifically because they felt OpenAI wasn't taking safety seriously enough. It was supposed to be the safety-first alternative. Now its own safety researchers are leaving because they don't think Anthropic is taking safety seriously enough. Competition as accelerant Coxon's warning specifically flagged international competition as a factor making things worse. The pressure to keep pace with rivals, particularly from Chinese AI labs, creates an environment where slowing down to get safety right feels like unilateral disarmament.
Anthropic projects 32% GDP growth by 2030 with AI scenarios
The Claude maker's most aggressive model envisions $44.4 trillion in US output, but only 10% of surveyed respondents actually buy it Anthropic just put a number on the AI hype train, and it's a big one. In the company's most aggressive economic scenario, US GDP could surge 32.4% above a no-AI baseline by 2030, hitting $44.4 trillion at 2025 price levels. That would mean annual growth rates touching 15% after 2027. Three roads diverge in an economic model The analysis lays out three distinct scenarios for how AI reshapes the American economy, each pegged to a different speed of adoption and deployment. The modest scenario projects just a 1.6% GDP uplift, bringing output to $34.1 trillion by 2030. The substantial scenario lands in the middle, projecting an 8.3% GDP increase that would push output to $36.3 trillion. This is where the typical respondent clusters. Most people surveyed expect something in the neighborhood of a 10% GDP uplift by 2030, paired with unemployment hovering around 5%. Then there's the extreme scenario. A 32.4% leap assumes rapid deployment of self-improving AI systems that automate and augment the vast majority of knowledge work. Not full replacement of human labor, but close enough that the productivity gains cascade through every sector of the economy. Growth stays relatively contained until AI diffusion hits critical mass after 2027, at which point the model projects sustained double-digit GDP expansion. The skepticism gap Perhaps the most revealing data point isn't about GDP at all. It's that only 10% of survey respondents aligned with the extreme growth projection. The majority of economists and industry respondents gravitate toward the substantial change scenario, expecting an 8-10% GDP increase alongside a slight rise in unemployment. What the models actually depend on The gap between $34.1 trillion and $44.4 trillion, between the modest and extreme scenarios, ultimately comes down to three variables: how fast AI systems improve, how quickly businesses deploy them, and how willing workers and institutions are to adapt. Anthropic's analysis draws partly on usage data from its own Claude models. The extreme scenario specifically envisions self-improving AI tackling knowledge work at scale, handling complex analysis, decision-making support, and creative problem-solving across industries. The model assumes these capabilities automate most tasks without fully replacing the humans overseeing them. The 5% unemployment rate that respondents associate with the substantial scenario suggests the consensus view is that AI creates enough new work to offset the jobs it automates, at least through 2030.
Anthropic warns Claude users of infostealer malware infections that hijacked active sessions
The AI company began contacting affected users after detecting unusual usage patterns tied to stolen browser cookies across multiple malware families. Anthropic started reaching out to impacted Claude users on August 30, 2026, after discovering that infostealer malware had compromised their computers and siphoned off active login session cookies. The stolen cookies gave attackers a backdoor into user accounts, letting them burn through paid usage quotas without ever needing a password or cracking two-factor authentication. What happened and how Anthropic responded The breach came to light through unusual usage patterns on Claude accounts. Once Anthropic's security team connected the dots, they moved quickly. All compromised sessions were forcibly signed out. Saved payment methods were stripped from affected accounts. And for users who got hit with unauthorized charges from attackers running up their quotas, Anthropic issued refunds. Anthropic was also careful to draw a clear line: the malware had nothing to do with Claude itself, its infrastructure, or anything users did on the platform. The infections originated from malicious downloads and compromised software applications that users had installed on their own machines. The malware families identified in the campaign included Vidar, LummaC2, StealC, RedLine, and Acreed on Windows. A limited number of Mac devices were also affected, primarily through Atomic Stealer, also known as AMOS. Why infostealers are targeting AI accounts The infostealer families involved in this campaign are well-established tools in the cybercrime ecosystem. RedLine has been one of the most widely distributed infostealers for years, while LummaC2 has surged in popularity among threat actors for its ability to harvest browser data, crypto wallet credentials, and session tokens. Vidar and StealC operate on similar principles, scraping stored credentials and cookies from browsers. Atomic Stealer has carved out a niche as one of the more capable macOS-focused threats. The common thread across all of these tools is their focus on browser session cookies. Modern web applications keep users logged in through session tokens stored as cookies. Steal the cookie, and you inherit the session. No password prompt, no 2FA challenge, no suspicious login alert. From the server's perspective, the attacker looks identical to the legitimate user. A broader pattern of AI-related threats This session-hijacking campaign is distinct from other security concerns that surfaced around AI platforms in 2026. Earlier in the year, separate reports documented attackers misusing Claude's own features or orchestrating phishing campaigns designed to deliver malware. Anthropic's latest advisory is specifically focused on the session-hijacking aspect rather than the methods of initial infection. Some security-conscious users have started adopting browser-level protections like Google Chrome's Device Bound Session Credentials, which ties session cookies to a specific device and makes them useless if exfiltrated.
Anthropic introduces local sandbox mode for Claude Code desktop
The new feature uses OS-level isolation to lock down command execution, cutting permission prompts by 84% since initial development began. Anthropic has rolled out a local Bash sandbox mode for Claude Code on desktop, giving the AI coding tool a security upgrade that isolates command execution at the operating system level. The feature works on macOS and Linux/WSL2, using native sandboxing technologies to restrict what Claude Code can actually touch on your machine. How the sandbox actually works The technical implementation varies by platform. On macOS, the sandbox relies on Seatbelt, Apple's built-in sandboxing framework that enforces fine-grained restrictions on process-level access. Linux and WSL2 users get bubblewrap, a lightweight containerization tool commonly used in the Linux ecosystem for unprivileged sandboxing. Both approaches accomplish the same goal: filesystem access gets locked down to the current working directory and its children, while network requests pass through a proxy layer that only permits connections to pre-approved domains. Windows users, for now, are left out. Full native support for Windows remains absent. The sandbox ships with two operational modes. The first is auto-allow, which lets commands execute without requiring explicit user approval each time. The second is a traditional permissions mode that still gates every command behind a manual check. Since Anthropic began iterating on this sandbox approach around October 2025, the company says it has achieved an 84% reduction in permission prompts. Security context and prompt injection defense The sandbox addresses a real and growing attack surface: prompt injection. A carefully crafted prompt injection could trick an AI assistant into running destructive commands, exfiltrating sensitive files, or establishing unauthorized network connections. The sandbox's filesystem restrictions and network allowlists serve as guardrails against exactly these scenarios. By confining execution to the working directory, even a successful prompt injection attack would struggle to reach SSH keys, environment variables, browser cookies, or other sensitive data stored elsewhere on the system. The network proxy adds a second layer, preventing exfiltration attempts to unauthorized domains. Anthropic's documentation makes clear that "computer use" features -- the desktop interaction capabilities that let Claude control mouse clicks and screen interactions -- run outside the sandbox environment. This means developers using those features still operate without the isolation protections the sandbox provides. The broader AI coding tool landscape Anthropic's sandbox development fits into a broader strategy that began taking shape in 2025, when the company started positioning Claude Code as a multi-functional development environment where AI can autonomously write, test, and execute code. The company also supports third-party sandboxing options, including Docker-based isolation, for users who want stronger separation between Claude Code's execution environment and their host system. Docker sandboxes offer a more comprehensive isolation layer than OS-level tools like Seatbelt or bubblewrap, though they come with additional setup overhead and resource consumption.
Anthropic quietly captured over 60% of business AI API spending, leaving OpenAI with roughly 35%
The enterprise AI race has flipped, with Anthropic commanding premium pricing on a fraction of the token volume while OpenAI fights to hold ground. The conventional wisdom that OpenAI dominates the AI market needs an asterisk. When you zoom into where businesses actually swipe their corporate cards for API access, Anthropic is running away with it, capturing north of 60% of spending while OpenAI sits at around 35%. That spending gap looks even more dramatic when you consider how it's being generated. According to Vercel AI Gateway data, Anthropic is pulling in that 61-65% spending share on just 30-32% of total token volume. That means Anthropic is charging roughly 4.4 times the average price per token compared to OpenAI. How the tables turned Rewind to 2023, and Anthropic held a modest 12% of enterprise LLM spend. OpenAI was sitting comfortably at around 50%. Fast forward to December 2025, and a Menlo Ventures survey showed Anthropic had climbed to 40% of enterprise LLM expenditures while OpenAI had slid to 27%. That trajectory only accelerated into 2026. Ramp's sales data from the same period showed Anthropic capturing 34.4-44% of business AI spending versus OpenAI's 32.3-40%. By mid-2026, the Vercel data tells an even starker story, with Anthropic pulling above 60%. The reversal happened fastest in one category: coding. Anthropic now holds approximately 54% of the enterprise coding and agentic applications market. OpenAI managed just 21% in that same segment. Average spending per user tells its own story. Anthropic users spend roughly $420 on average, compared to $310 for OpenAI users. Different strategies, different moats Anthropic's enterprise and API revenue mix stands at approximately 80%, meaning the vast majority of its money comes from businesses integrating Claude into their products and workflows. OpenAI, by contrast, has built a large consumer product in ChatGPT. Its consumer-scale reach gives it a different kind of leverage: volume. OpenAI's broader product ecosystem, from ChatGPT Enterprise to its Azure partnership, gives it distribution channels that pure API metrics don't capture. The multi-model reality One data point complicates any simple narrative about winners and losers: 67% of development teams now use multiple AI providers. Teams route their hardest, most valuable work to Claude -- the complex coding tasks, the multi-step agentic workflows -- while simpler, high-volume tasks get sent to cheaper alternatives. Some Ramp data showed stronger growth for OpenAI in certain business cohorts during Q3 2026, suggesting the company hasn't ceded the enterprise fight entirely.
Anthropic upgrades Claude's streaming renderer, claims 9x fewer stalls on slower laptops
The frontend optimization delivers smoother token streaming for long responses without changing Claude's underlying models or API Anthropic rolled out a frontend optimization to Claude's streaming response renderer on August 24, targeting a problem that anyone who's tried to get a long answer from an AI chatbot on an older machine knows well: the maddening stutter-and-freeze cycle that turns a conversation into a slideshow. The update applies to Claude's web and desktop applications. According to Anthropic, long answers now stream roughly 4x smoother, stalls on slower laptops drop by 9x, and worst-case freezes shrink by 4.5x. On hardware that supports it, like 120 Hz MacBooks, the renderer can sustain 120 fps during streaming output. What actually changed under the hood The core fix is elegant in its simplicity. Previously, each incoming token from Claude's model triggered a repaint of the full response container, or at least a substantial portion of it. The new approach restricts UI updates to only the elements that have actually changed, resulting in dramatically less computational overhead per token, which matters most on machines with limited GPU headroom or older processors. Critically, Anthropic made no changes to its models or API with this release. Generation speed, output quality, and the underlying inference pipeline remain untouched. This is purely a presentation-layer improvement. Why UI performance matters in the AI race The 120 fps sustained frame rate figure is particularly telling. Most web applications don't need to think about frame rates at all, but streaming text renderers are effectively animations. Each new token is a frame update. When you're generating hundreds of tokens for a long response, that's hundreds of sequential frame updates, and any hitch becomes visible as a stutter or freeze. Hitting 120 fps on compatible hardware means Anthropic's renderer can now match the refresh rate of Apple's ProMotion displays without dropping frames. The feedback loop, and its limits User reception on social media was largely positive, with many noting that the improvement was immediately noticeable during extended conversations. But some responses highlighted a tension that Anthropic and every other AI company faces: polishing the interface doesn't fix the model. Several users pointed out that smoother streaming doesn't help when Claude hallucinates a citation or confidently delivers incorrect information. The update also has implications for Anthropic's enterprise ambitions. Corporate users often run standardized hardware that skews older than what developers and early adopters use. A 9x reduction in stalls on slower laptops isn't just a nice stat for a blog post. It's the difference between an enterprise deployment that employees actually use and one they abandon for a competitor after a week of frustrating freezes.
David Sacks accuses Anthropic of regulatory capture against open source AI
The White House tech advisor warns that a proposed self-regulatory body could quietly kill open source AI models through compliance requirements they can never meet David Sacks, venture capitalist and co-chair of the President's Council of Advisors on Science and Technology, is sounding the alarm on what he calls a quiet regulatory strategy to suffocate open source AI. His argument: you don't need to ban something outright if you can just regulate it into irrelevance. During an episode of the All-In Podcast, Sacks laid out a scenario in which a self-regulatory organization designed for AI oversight gradually morphs into a mandatory pre-release approval agency. Think of it as a DMV for AI models, except instead of waiting in line to renew your license, developers would need to get their models blessed before releasing them to the public. The compliance trap The core of Sacks' concern is structural. A pre-release approval regime would impose compliance requirements that proprietary, closed models from companies like Anthropic could feasibly meet. Open source models, by their very nature, cannot. Once you release an open source model into the wild, it's out there. It's decentralized, forkable, and immutable. You can't recall it for a safety audit the way Anthropic can update Claude behind a closed API. Requiring pre-release certification would essentially create a regulatory framework where closed models pass and open models fail by default. Sacks specifically named Anthropic, the company led by CEO Dario Amodei, as the primary actor pursuing what he described as "sophisticated regulatory capture." In his telling, Anthropic has positioned itself as the responsible adult in the room, advocating loudly for AI safety while quietly lobbying for regulatory structures that happen to favor its own business model. The industry split What makes Sacks' framing notable is the degree of isolation he attributes to Anthropic. According to him, virtually the entire tech industry supports open source AI development, with Anthropic standing as the notable exception. Meta, which has invested heavily in its open-weight Llama model family, represents the other end of the spectrum, betting that open release accelerates adoption and ecosystem development. The competitive dimension extends well beyond Silicon Valley. Chinese-developed open-weight models have recently topped or approached critical benchmarks, a development that Sacks has used to sharpen his argument about American competitiveness. If the US constrains open source AI through regulatory friction while China faces no such limitations, the talent and innovation gap could widen in the wrong direction. Sacks has been vocal on X about this framing, repeatedly invoking the principle of "permissionless innovation," the idea that developers should be able to build and release technology without needing prior government approval. What this means for the AI landscape The companies most exposed to this regulatory risk are the ones building in the open. Meta's Llama ecosystem, Mistral, Stability AI, and the broader constellation of startups and research labs that depend on freely available model weights all face a scenario where their core distribution model becomes legally complicated, if not outright impossible. Sacks' warning also carries weight because of his current position. As co-chair of the President's Council of Advisors on Science and Technology, he's not just a podcast commentator. He has a direct channel to policy discussions, which means his framing of the issue, regulatory capture dressed up as safety, could influence how the White House approaches AI governance.
Vercel reports open-weight models hit 62% of AI Gateway traffic in August
Open-weight models nearly tripled their share of Vercel's token volume in two months, reshaping how enterprises think about AI spending Two months ago, open-weight AI models were a minority player on Vercel's infrastructure. Now they're running the show, at least by volume. Vercel CEO Guillermo Rauch reported on August 22 that open-weight models accounted for 62% of all tokens processed through Vercel's AI Gateway, up from 28.4% on June 24. How fast is fast? Open-weight models held just 11% of Vercel's token volume in April. By June they were at 29%. By late August they crossed 62%. Vercel's AI Gateway acts as a routing and traffic management layer for AI-powered applications, meaning its data reflects real production workloads from real companies, not benchmark experiments or lab conditions. The structural thing happening here is cost. Open-weight models can run at roughly one-tenth the price of their closed-source counterparts, and enterprises have figured out that not every AI task needs a premium model to get the job done. The spending paradox Despite commanding 62% of token volume, open-weight models are not commanding 62% of the money. Anthropic's closed models captured between 61% and 65% of total expenditure on the gateway in recent reporting periods. Claude is processing a minority of tokens but collecting a majority of the revenue. DeepSeek has climbed to the top, or near the top, of Vercel's token volume leaderboard, overtaking Google in processing share. What this means for AI deployment broadly Major companies including AT&T and Coinbase have been noted among those emphasizing cost-reduction strategies in AI deployment, which aligns with exactly the kind of workload routing the Vercel data describes.
